acceptodds
Withdrawn from ICLR 2027

MixORNet: Encoding Scent Mixtures via Combinatorial Protein Activations for Olfactory Perception

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Perceptual modeling of complex scent mixtures remains largely underexplored in computational olfaction. Existing approaches share two critical limitations: (1) they rely solely on the chemical properties of constituent odourants, neglecting selective olfactory receptor (OR) binding, and (2) they model perception via pairwise molecular attention, ignoring the fact that humans perceive mixtures holistically. We propose MixORNet, the first framework to encode scent mixture perception at the receptor level. MixORNet represents the mixture as a fully connected graph and uses a graph attention network (GAT) to extract holistic mixture-level features. These features further interact with a panel of OR proteins via cross-attention, yielding -dimensional OR-activation fingerprints, with each dimension denoting the activation probability of each receptor. MixORNet achieves state-of-the-art performance on perceptual distance prediction and multilabel olfactory annotation of scent mixtures. Moreover, it can reproduce the olfactory white phenomenon, aligning with established psychophysics. Notably, ablation of the OR panel size reveals that the optimal is substantially smaller than the approximately 400 human ORs in specific scenarios, providing the first AI-derived estimate for the intrinsic dimension of the human olfactory perceptual subspace. The main code is available in the anonymous repository https://anonymous.4open.science/r/MixORNet/

Primary area:
applications to neuroscience & cognitive science
Keywords:
Computational Olfaction, Scent Mixture, Olfactory Reactor, Molecule-protein interactions, Chemical multimodal framework

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_aB4COpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MixORNet, which combines representations of the molecules in a mixture with ESM-2 features of olfactory receptor proteins through cross-attention to predict perceptual attributes of mixtures and perceptual distances between them. It achieves competitive performance against closely related methods on two DREAM benchmarks and reports higher mean performance on the unseen-molecule splits. Its main contribution is the integration of receptor information into mixture representations and an empirical evaluation under multiple conditions.

Critical strengths

S1. A coherent approach to an important and timely problem Predicting perceptual attributes and distances at the mixture level is an important problem, given that real-world odors commonly consist of multiple molecules. Extending the use of pretrained molecular and protein representations from single molecules to mixtures is a timely direction for olfactory modeling, with potential applications in fragrance design and olfactory sensing. The architecture provides a concrete approach by connecting permutation-invariant mixture representations with receptor-conditioned representations.

S2. Relevant comparisons and competitive empirical results The paper evaluates strong, relevant methods, including POMMix and MixScentNet, on the same data splits and considers both standard and unseen-molecule settings. MixORNet is competitive with leading methods on the standard splits and reports higher mean performance on both tasks under the unseen-molecule protocol. For example, on the unseen DREAM2024 split, PLCC is 0.6724 for MixScentNet and 0.7045 for MixORNet. These results, together with the component analyses, support the practical usefulness of the approach and are the main grounds for my positive assessment.

Critical weaknesses

W1. It is unclear whether the unseen-molecule split excludes all test constituents from training. Section 4.2 states that no SMILES appearing in a test sample occurs in a training sample. However, assigning all samples containing a selected molecule to the test set does not by itself establish whether the other constituents of those test mixtures are also excluded from training. Holding out selected molecules and excluding every test constituent from training support different interpretations of generalization. Since the results in Table 2 are a central part of the empirical contribution, the handling of co-occurring constituents and the actual train/test molecule-overlap statistics are important for assessing these results.

W2. The current experiments do not isolate the contribution of receptor-sequence information to the generalization gains. Section 4.2 attributes the gains on unseen molecules to the OR fingerprint and receptor-based regularization. However, the comparisons between complete models in Tables 1 and 2 do not distinguish the effect of receptor-sequence content from that of the additional conditional representation architecture. The ablation in Table A1 that removes cross-attention still retains the protein features and is reported only on the standard split. A control that changes only the receptor input features while preserving the architecture would help address this question, particularly on the unseen-molecule splits where the main gains occur. The existing results support the predictive performance of the complete model more directly than the proposed explanation for its improvement.

Questions for authors

Q1. Does the unseen-molecule construction guarantee that every constituent of a test mixture is absent from training, or is this guarantee limited to the molecules selected for holdout? Please explain how the other constituents that co-occur with a selected molecule are handled, and report the train/test sample counts and the size of the intersection between the constituent-molecule sets for each fold. For DREAM2024, please clarify whether constituents from both mixtures in each pair are included when checking overlap, and whether the five test folds are mutually exclusive partitions or independently sampled holdouts. Fold-wise paired performance differences against MixScentNet would also help assess the consistency of the reported gains.

Q2. How much of the improvement depends on the sequence-derived receptor features? Could the authors provide a control that preserves the receptor panel and architecture but replaces the receptor features with random features drawn once for each receptor and held fixed, while retraining the remaining trainable modules? Matching input dimensions, residue counts, and feature magnitudes, and using the same splits and training configuration, would help assess the additional benefit of the actual receptor features. This comparison would be particularly informative on the unseen-molecule splits. If such results are unavailable, please clarify how strongly the current evidence supports attributing the improvement to receptor features.

Q3. For each fold in Tables 1 and 2, were model-selection decisions made without using that fold's test results? Section 3.4 motivates N=100N = 100 by performance and efficiency. Please specify which data and criteria were used to select the panel size, the stopping epoch or checkpoint, and TrampT_{\mathrm{ramp}}. Please clarify whether these settings were fixed in advance, selected using a separate validation procedure within each fold, or chosen based on the reported cross-validation results. Describing the validation and model-selection procedures used for the baselines would also help assess the comparability of the results.

Reviewer_J93iOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces MixORNet, an attention-based approach to predict mixture odor perception using per-stimulus olfactory receptor (OR) fingerprints derived from components' molecular graphs and receptor protein sequences. The paper shows that this approach achieves comparable or better performance to the evaluated benchmarks in two split types (standard cross-validation and unseen components) on two datasets from DREAM olfactory challenges organized in 2024 and 2025. The model further predicts lower perceptual distance to mixtures with the increasing number of components and reports performance saturation with substantially fewer ORs than the full human repertoire (roughly 40-100 ORs depending on the setting).

Critical strengths
  • Interesting biological inductive bias. The paper introduces a receptor-conditioned architecture for mixture perception, using OR sequence representations to construct a biologically motivated intermediate representation rather than relying only on molecular or perceptual embeddings. This is a meaningful modeling direction even if the broader novelty claim should be narrowed.

  • Promising generalization results. The strongest empirical result is the unseen-molecule evaluation, where MixORNet improves over MixScentNet on both DREAM2024 and DREAM2025. This suggests that the receptor-conditioned representation may provide useful inductive bias for generalization beyond molecules observed during training.

  • Reasonably thorough architectural analysis. The ablation experiments show that both the mixture graph-attention component and the receptor cross-attention mechanism contribute to performance, which helps support the architectural design choices.

  • Scientifically motivated evaluation beyond benchmark scores. The paper attempts to connect the learned representation to psychophysical phenomena such as olfactory white and to study performance as a function of the number of receptors. Even though I think some of these interpretations are overstated, the effort to test biologically meaningful hypotheses rather than only report leaderboard metrics is a strength.

  • Clear overall presentation and reproducibility effort. The architecture, losses, receptor-panel construction, and training setup are described in detail, and the paper reports multiple datasets, baselines, ablations, and appendix analyses.

Critical weaknesses
  • At several places in the paper, the authors claim that the resulting OR fingerprint is biologically interpretable as OR-activation probabilities (e.g., line 91, 132, 369). However, this claim is never tested in the paper, and it is possible that the ORs representations work as fixed abstract anchors similar to the learned inducing points in Set Transformer (Lee et al., 2019) rather than encoding biologically meaningful receptor activation.

  • Incomplete comparison against state-of-the-art DREAM methods. The paper's central empirical claim is state-of-the-art performance on DREAM2024 and DREAM2025, but the winning models and their ensemble are not evaluated in Tables 1 and 2. Instead, they appear only in an appendix leaderboard comparison, where results were obtained under the competition evaluation rather than by rerunning these methods under MixORNet's evaluation protocol. This is insufficient to establish that the proposed method improves over the current SOTA.

  • The authors claim that "MixORNet is the first model to introduce receptor-level combinatorial coding into scent mixture perception, which is a paradigm shift from existing QSOR models" (line 87). However, in the scope of both DREAM challenges there are prior models that already incorporate olfactory receptor activation: ChemSenSimLab in DREAM 2024 (Satarifard et al., 2026) and Nylix model in DREAM 2025 (see DREAM 2025 write-up and https://github.com/Maxence-Lalis/NyDream2025). Both these models appear in Table A2 but are not mentioned in the Related work section nor evaluated in Table 1 and Table 2.

Questions for authors
  • Can the authors provide evidence that the OR fingerprints encode receptor-specific biological information rather than merely providing useful conditioning variables? This can be possibly tested from several different angles (see below), where the first experiment is inference-only:

      1. Test whether OR fingerprints predict the response of in vitro experiments on single molecule-OR pairs available in M2OR. Even though MixORNet is designed for mixtures, it should at minimum exhibit meaningful correspondence to known interactions between single molecules and ORs. As noted by Hladis et al., 2023, this evaluation should only consider EC50 experiments, as screening can introduce label noise.
      1. Secondly, a possible explanation of the w/o CA ablation in Table A1 is that OR sequences are used by the model as abstract latent space anchors, similarly to inducing points in the ISAB module of Set Transformer (Lee et al., 2019). Therefore, the authors should evaluate the performance of MixORNet when fixed random embeddings matched in dimensionality and scale to the projected ESM-2 OR embeddings are used instead of ESM-2 embeddings encoding OR sequences. If time and compute permits, an extension is to evaluate learned embeddings similar to ISAB in Set Transformer, but I would understand if this is too demanding given the time for the rebuttal.
      1. Finally, evaluate the performance gain from using OR sequences compared to ESM-2 embedding of protein sequences unrelated to olfaction. In particular, what is MixORNet's performance when ORs are replaced by the same number of protein sequences from GPCR class A family?
  • The above question also relates to the estimated number of ORs used in biological odor perception. Does performance saturate similarly to Figure 3 if the OR embeddings are replaced by random vectors or unrelated protein embeddings?

  • Please evaluate the strongest reproducible DREAM2024/2025 baselines under the same splits. In particular, specifically evaluate the models that use OR activation as an input (ChemSenSimLab in DREAM 2024 and Nylix in DREAM 2025) or their closest equivalent that can be implemented based on the open-source code (EC50 model in Nylix may not be available, but it can be replaced by the OR-activation model introduced by Hladis et al, which is used in the ChemSenSimLab model). In addition, currently reported comparisons with other approaches in DREAM 2024 may be confounded by the task-specific MSE-to-PLCC loss used by MixORNet. The loss ablation shows that PLCC optimization accounts for a large fraction of DREAM2024 performance (Table A1), so strong baselines should ideally be evaluated with the same objective.

  • The DREAM2024 results appear to depend substantially on the choice of training objective. In Table A1, replacing the default MSE-to-PLCC loss with MSE reduces PLCC from 0.7842 to 0.7193, whereas PLCC loss alone already reaches 0.7833. Thus, it is unclear how much of the reported improvement over prior methods is attributable to the proposed architecture versus directly optimizing the evaluation metric. Were the competing neural baselines trained with the same MSE-to-PLCC objective? If not, could the authors retrain the strongest baselines with this loss, or otherwise disentangle the gain from the architecture from the gain due to metric-aligned optimization?

  • Considering the olfactory white experiment, the authors claim that the odor similarity tends towards zero faster for MixORNet then for human subjects. Why should this discrepancy be interpreted as recovery of latent psychophysical structure rather than an overly strong mixture-complexity bias learned from the data? In addition, for the Ravia olfactory white analysis, are the reported human correlations computed after pooling Ravia 1 and Ravia 2 with Ravia 4? These three experiments use different protocols (Ravia 1 and 2 use direct similarity ratings and Ravia 4 is triangle test). If the results are pooled, how do the correlations look when computed separately by experiment or after accounting for experiment type? Because these paradigms have different floors (for triangle test, chance performance is 1/3, so values near or below 1/3 do not represent progressively smaller perceptual distances in the same sense as values near zero on the direct similarity-rating scale) and psychometric scales, pooling them may attenuate the human complexity correlation and make the model-vs-human discrepancy appear larger than it is.

Reviewer_oDe4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MixORNet, which encodes an odor mixture with an attention network and then matches it against a panel of N olfactory receptors, represented by frozen ESM-2 embeddings, to produce an N-dimensional "OR-activation fingerprint". The fingerprint is used to predict odor descriptors (DREAM2025) and perceptual distances (DREAM2024), and the receptor module is trained only through these perceptual losses. The method is comparable to MixScentNet on random splits and better on unseen molecules. The authors also claim it reproduces olfactory white and that 40–100 receptors are enough.

Critical strengths

1 I like the idea of using a receptor panel as an intermediate representation is an interesting, biologically motivated idea.

2 Receptors enter only through their sequences and share all weights, so the model could in principle score unseen receptors or OR variants.

3 Gains on the unseen-molecule split are clear, especially on DREAM2025, where the loss is plain MSE.

Critical weaknesses

1 There is no grounded evidence shows that the "OR fingerprint" relates to real receptors. No receptor-level supervision is used, and since all receptors share the same weights, the ESM-2 embeddings may simply act as arbitrary anchors. The authors should compare against random anchors, non-olfactory proteins or pseudogenes, and learnable embeddings, and check the outputs against M2OR activations.

2 The gains could possibly come from the loss rather than the receptor module. With plain MSE, MixORNet reaches only 0.719 PLCC on DREAM2024, below POMMix (0.756) and MixScentNet (0.785). The jump to 0.784 comes from the PLCC loss, and PLCC alone already gives 0.783, so the gain amounts to optimizing the evaluation metric directly. The baselines should be retrained with the same loss and distance.

  1. The intrinsic-dimension claim does not follow. Saturation at 40–100 receptors more likely reflects the bottleneck width that two small datasets need than anything about human olfaction.

Minor: The unseen-molecule split only excludes the seed molecules. Other components of test mixtures can still appear in training, so the claim that no test SMILES appears in training is not supported.

Questions for authors

See my major concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.